GLM-4.6V SFT LoRA (T1plus)

Fine-tuned LoRA adapter for GLM-4.6V 108B MoE Vision-Language Model.

Model Details

• Base Model: GLM-4.6V 108B MoE (128 experts, 8 active)
• Training Method: SFT with LoRA
• LoRA Rank: 64
• LoRA Alpha: 128
• Training Epochs: 2
• Learning Rate: 2e-05
• Max Sequence Length: 4096

Training Configuration

• Batch Size: 1
• Gradient Accumulation: 8
• Precision: bfloat16

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "THUDM/GLM-4.6V",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "HUNGTZE/T1plus")

Framework Versions

• PEFT 0.18.0
• Transformers 4.x